OpenAI will begin adding invisible watermarks to text responses from ChatGPT and Codex in the European Union in the coming weeks. This was reported by Qazaqyia.kz citing Kursiv Media.
The marking will help determine whether the company's tools were used to create a text, according to an OpenAI statement.
First stage of rollout
In the first stage, the innovation will affect users of all tariff plans in the EU, but will apply only to texts suitable for marking. The company does not yet plan to introduce it by default worldwide. API clients — the interface for connecting models to third-party services — can already voluntarily enable watermarks for individual models regardless of country.
textGrain technology
OpenAI's technology called textGrain imperceptibly adjusts the choice of words and their parts when generating a response. As a result, a statistical pattern arises in the text that a special detector can recognize.
For the reader, the watermark is invisible: the system does not add hidden characters, unusual spaces or punctuation marks. The marking is built into the word selection itself and can persist when copying text. According to the company, its impact on model performance is negligible, and tests did not reveal a significant deterioration in response quality.
Access to the detector
At launch, the detector will not be publicly available. Access will be provided to approved researchers and expert organizations who will help assess the reliability of the technology.
At the same time, a detected watermark does not reveal the user's identity, their requests or correspondence. It also does not show what contribution a person made to preparing the text, and does not confirm the accuracy of the information.
Limitations of the technology
OpenAI acknowledges the limitations of the technology. Short texts and responses with a small choice of formulations are harder to check. Substantial editing, retelling or translation can weaken the marking and make it invisible to the detector.
Therefore, the absence of a watermark does not prove that a human wrote the text. The system may also falsely detect a marking where there is none.
Compliance with EU law
OpenAI explained the introduction by the need to comply with the European law on artificial intelligence — the EU AI Act, which provides for machine-readable marking of generated content.